The E ect of Education Policy on Crime: An Intergenerational
Transcrição
The E ect of Education Policy on Crime: An Intergenerational
The Eect of Education Policy on Crime: An Intergenerational Perspective∗ † ‡ § Costas Meghir , Mårten Palme , Marieke Schnabel May 2011 Abstract The Swedish comprehensive school reform implied an extension of the number of years of compulsory school from 7 or 8 to 9 for the entire nation and was implemented as a social experiment by municipality between 1949 and 1962. A previous study (Meghir and Palme, 2005) has shown that this reform signicantly increased the number of years of schooling as well as labor earnings of the children who went through the post reform school system, in particular for individuals originating from homes with low educated fathers. This study estimates the impact of the reform on criminal behavior: both within the generation directly aected by the reform as well as their children. We use census data on all born in Sweden between 1945 and 1955 and all their children merged with individual register data on all convictions between 1981 and 2008. We nd that the educational reform decreased crime substantially for men who were directly aected by it. We also nd that the crime rate declined for the sons of those fathers directly aected by the new educational system; we interpret this results as implying that improved education increased resource and parenting quality, leading to improved child outcomes. We thank Matthew Lindquist, Imran Rasul, Lisa Jönsson, Emilia Simeonova as well as participants at seminars at Stockholm University, University College London, Yale University and the Institute for Labor Market Policy Evaluation (IFAU) in Uppsala for helpful comments and suggestions. We gratefully acknowledge nancial support from IFAU. Costas Meghir thanks the ESRC for funding under the Professorial Fellowship RES-051-27-0204 and under the ESRC Centre at the IFS ESRC RES-544-28-5001. † Department of Economics, Yale University; Institute for Fiscal Studies and Department of Economics, University College London, Gower Street, London WC1E 6BT, UK. E-mail: [email protected] ‡ Department of Economics, Stockholm University, SE - 10691 Stockholm, Sweden. E-mail: [email protected] § Department of Economics, University College London, Gower Street, London WC1E 6BT, UK. E-mail: [email protected] ∗ 1 1 Introduction Crime imposes huge costs on society and policies to prevent crime are debated around the world. In addition to police resources and punishment of criminals, these policies typically involve public spending on creating alternative labor outcomes for individuals potentially involved in criminal activities, e.g. reforms of national education systems. Since criminal behavior has an exceptionally strong intergenerational link, such policies may have an eect across generations. Although there is a very large previous literature showing that low educated people are more likely to be convicted and that their children are more likely to commit crimes, very few studies have attempted to show a causal relation between own education and criminal behavior and (to our knowledge) no previous study has shown a causal relation between parental education and criminal behavior. In this paper we study to what extent a major national education reform, which increased schooling and earnings, also aected criminal behavior: Both in the generation directly aected by the reform as well as their children. The reform in question is the Swedish comprehensive school reform, which included an extension of the number of years of compulsory schooling and was implemented gradually across the 1,055 Swedish municipalities during the 1950s, before nally being rolled out nationally in 1962. As described in Meghir and Palme (2005), the implementation of the reform allows us to investigate its impact on crime by comparing individuals living and working in the same labour market but having been exposed to dierent schooling systems. Moreover, we can also estimate the impact of the reform on their own children. In this case we are able to consider how changing the educational experience of the parents can aect important outcomes for their children. Beyond the obvious public policy implications, the results speak to the growing literature 2 on the importance of parenting quality on child outcomes. Two earlier papers by Lochner and Moretti (2004) and Machin, Marie, and Vuji¢ (2010) respectively study the relation between compulsory schooling laws and criminal behavior. Lochner and Moretti (2004) use changes in compulsory schooling laws across time between US states to identify the eect of increasing education on crime. Machin, Marie, and Vuji¢ (2010) compare criminal behavior of the cohorts just before and just after the implementation of the comprehensive school system in Britain. Our paper goes further in two important ways. First, we compare the criminal behavior of two groups, distinguished by the school system they were exposed to, but active in the same labor markets at the same point in time, and who belong to the same cohort. As a result any general equilibrium eects of the reforms are eliminated, as they are common across groups - what we identify is the impact of being exposed to alternative educational systems, over and above the macroeconomic eects that result from this. However, the key distinguishing feature of our paper is that we consider the eects on the criminal activity of the children of those aected by the reform. As we discuss later, the eects in this case will operate through improved parenting and investments in children. The paper is organized as follows: Section 2 discusses previous theoretical and empirical work on the relation between both own education and criminal behavior as well as parental education and criminal behavior; Section 3 provides an overview of the comprehensive school reform in Sweden; Section 4 describes the data and presents empirical results on the relation between educational attainments and criminal behavior as well as intergenerational associations of crime; Section 5 discusses our identication strategy; Section 6 presents our empirical results on the eect of the comprehensive school reform and the probability of being convicted for crime or 3 sentenced to prison; Section 7 concludes. 2 The Impact of Education on Crime 2.1 The Impact of Education on Crime within a Generation The links between economic incentives and crime have been established both theoretically and empirically in earlier studies. A prominent example is Freeman (1999) who outlines an economic model of crime where the choice between criminal and legal activity is determined by comparing the expected utility of each. Grogger (1998), Gould, Weinberg, and Mustard (2002) and Machin and Meghir (2004) demonstrate the importance of wages and labour market opportunities in driving crime. One implication of this is that improved education may reduce crime. A number of papers have looked at the link between education and crime directly. These include Lochner and Moretti (2004) and Machin, Marie, and Vuji¢ (2010), cited above. A more theoretically based approach was oered by Lochner (2004) who develops a life cycle model of education and crime and estimates a negative education-crime relationship. A study, based on this capital approach by Williams and Sickles (2002) nds that years of schooling has a signicant negative eect on crime in adulthood, and that there is a relationship between crime and other measures of human capital. Earlier studies support this empirical evidence on the education-crime relationship; for example Freeman (1996) states for the 1991 US Census that two thirds of US prison inmates are high-school drop-outs and 12 percent of 24-35 year old high school drop outs were incarcerated in 1993. This negative correlation between crime and education has also been documented in the the crim- 1 inology and sociology literature, for example Sabates and Feinstein (2008a). 1 See also Sabates and Feinstein (2008b). 4 Gallipoli and Fella (2006) estimate a general equilibrium model of crime and education. They nd that increases in education have a signicant impact in reducing crime. However, they point out that the general equilibrium eects, operating through changes in wages as the number of educated individuals increases, can be substantial. 2.2 The Eects of Parental Education on Children's Criminal Behavior Intergenerational associations of criminal behavior have been documented in the criminology literature. From the economics point of view this question relates to the investments that parents make on their children and the way that parental education may aect such investments (see Becker (1981)). 2 Of particular interest to our study are the two recent papers by Hjalmarsson and Lindquist (2010a) and Hjalmarsson and Lindquist (2010b). The rst documents a strong correlation between crime of fathers and children of both genders in Sweden using the Stockholm Birth Cohort Study, which is a data set that consists of all individuals born in 1953 and living in Stockholm in November 1963, containing crime records of the birth cohort and their fathers. The second study focuses on parent-child correlations in crime using adoption data. In particular they aim to determine through which factors mothers and fathers inuence child criminality, which is closely related to the study by Björklund, Lindahl, and Plug (2006). 2 For some empirical work see for example Carneiro, Meghir, and Parey (2007). Moreover, there is direct evidence that better childhood environments and early education can reduce crime rates, see for example the results form the Perry pre-school experiment presented in Schweinhart, Montie, Xiang, Barnett, Beleld, and Nores (2005) and Cunha and Heckman (2007). 5 2.3 Human Capital and participation in Crime - a simple framework To better understand the mechanisms through which educational reform can aect participation in crime for both generations consider the following simple model. Human capital is produced by investments in various stages of the child's life as well as by overall educational attainment investments I0 Ec . Suppose there are two stages, early and investments during schooling depends on the educational level of the parent, I1 . The eciency of investments Ep . Denote the human capital production function by H = H(I0 , I1 , Ec |Ep ) where HI0 0 > 0, HI0 1 > 0, HE0 c > 0 and HI000 < 0, HI001 < 0, HE00 c < 0. Parents are assumed to care about child quality, which here is just their human capital. Ignoring dynamics for simplicity, they solve the problem max {u(C, H) C,I0 ,I1 where C st C + I0 + I1 = Y p is parental consumption. will increase as parental resources 3 and H = H(I0 , I1 , Ec |Ep )} In this simple context investments in children YP increase, so long as H is a normal good. The rst order conditions for investments are ∂H = u0C u0H ∂I 0 ∂H u0H ∂I = u0C 1 An increase in the marginal productivity of such investments (say due to an increase 3 The here. problem is dynamic sequential, but nothing would be gained in introducing this notation 6 in parental education Ep ) will lead to more investments in the children at both stages. This will happen both because the productivity of investments may increase and because parental resources YP go up. The next step is to see how these changes can aect participation in crime. Consider a very simple model of crime choice in two stages of life. First, is the educational stage, where the individual can either engage in education, crime or work. Then follows the adult stage where there is no education choice. We start by the latter. Committing a crime is a period by period decision with no dynamics (for simplicity). Working leads to income Y (H) which depends on human capital H . on the other hand yields a return with probability 1 − p(H) R with some probability leads to punishment Moretti (2004), the dependence of K on ing incarcerated. Moreover, it also costs H K(H). p(H). Being caught, As discussed in Lochner and represents the opportunity cost of be- c(H) to participate. This cost can reect the aversion that one may have to anti-social behaviour. We assume that increasing in Crime c(H) is H. Participation in crime is determined by the condition p(H)R − (1 − p(H))K(H) − c(H) − Y (H) > 0 ⇐⇒ An increase in H will increase earnings implying a reduction in crime. engage in crime Y (H) and participation costs c(H), both A possible mitigating eect is that better human capital may make crime more eective and reduce the probability of capture p(H). In our empirical analysis we only measure convictions; we assume that a reduction in convictions reects a reduction in crime participation and not more eective 7 criminals. Thus, other than the potential eect on p(H), increasing human capital will decrease participation in crime. It is also useful to consider the earlier period, when the individual still has the option of being in school. Dene the future value as with Vκ denoting the value of crime and Vw V (H1 ) = E maxκ,w (V κ (H1 ), V w (H1 )) the value for work. the cost of education, which we assume are declining in H0 ced (H0 ) denotes (initial human capital). In this rst period the value of education, crime and work respectively are given by V ed = −ced (H0 ) + βV (H1+ ) V w = Y (H0 ) + βV (H0 ) V k = p(H0 )R − (1 − p(H0 ))K(H0 ) − c(H0 ) + βV (H0 ) where H1+ denotes that education allows the individual to enter the next period with higher human capital. The individual choose the activity with the greatest value. First, note that if schooling is compulsory, then there is a mechanical reduction in crime, simply because the opportunity to commit an oence is no longer there (or reduced in practice). Second, an increase in human capital will increase the value of both schooling and work; the former because it will reduce the costs of schooling ced (H0 ) as well as the future value V (H1+ ), current wage as well as the future value the latter because it will increase the V (H0 ). So rst period crime will decline; whether education will go up is in this context ambiguous. We can now summarise the way that educational reform and compulsory schooling in particular can aect crime across generations. Meghir and Palme (2005) document the impact of the reform on educational attainment and earnings. 8 For understanding the impact of the reform they emphasise the importance of distinguishing individuals aected by the reform with parents with just the statutory level of education (60% in that population) to those whose parents had attained a higher level. An increase in compulsory schooling reduces the available time for crime early on; it increases human capital and thus reduces further the incentive to commit crimes and may increase the chance of remaining in school beyond the new compulsory level. H. It may also draw increased investments I1 from parents further increasing This reduces crime in the young (school period) ages. As an adult, the result is increased human capital, which as explained before, will reduce adult crime. If there is a habit formation aspect of crime (not included above), the early decline will be reinforced. Thus crime will decline relative to the group that was not aected by the reform. The children of the aected generation are all experiencing the same education system because the reform was rolled out nationally in 1962. They dier by the fact that some have parents who faced the new education system and as a result have more parental education and more resources to higher I0 , I1 , Y p. These dierences will lead and eventually higher educational attainment Ec relative to the children in the comparison group, whose parents did not go through the reform. Educational attainment Ec may increase because, according to mounting evidence, an increase in early investments I0 improves cognition and social skills and hence reduces the costs of education. In addition the increased resources Yp allow more transfers to children alleviating somewhat nancial constraints for education. These channels imply an increase in both H0 and described above. 9 H1 reducing crimes at all life stages, as For individuals aected by the reform but having parents with more than statutory education the impact is less clear cut. For this group there is no eect on educational attainment. However it changed the way they were educated because it abolished early selection and tracking, which aected primarily this group. It can be argued that quality of education for this group was diluted for this reason and because the increase in compulsory schooling, aecting the other group, could have reduced the quality of the peers. For this reason we cannot be condent that H increased for this group. This is why in our empirical analysis we present overall results as well as results separately on the lower socio-economic group. 3 The 1950 Education Reform and the Social Experiment 3.1 The Swedish school system before and after the reform Prior to the implementation of the comprehensive school reform, pupils attended a common basic compulsory school ( folkskolan ) until grade six. After the sixth grade pupils were selected to either continue one or, in mainly urban areas, two years in the basic compulsory school, or to attend the three year junior secondary school ( realskolan ). The selection of pupils into the two dierent school tracks was based on their past performance, measured by grades. The pre-reform compulsory school was in most cases administered at the municipality level. The junior secondary school was a prerequisite for the subsequent upper secondary school, which was itself required for higher education. In 1948 a parliamentary school committee proposed a school reform that im- 10 4 plemented a new nine-year compulsory comprehensive school. The comprehensive school reform had three main elements: 1. An extension of the number of years of compulsory schooling to 9 years in the entire country. 2. Abolition of early selection. Although pupils in the comprehensive schools were able to choose between three tracks after the sixth grade - one track including vocational training, a general track, and an academic level preparing for later upper secondary school - they were kept in common schools and classes until the ninth grade. 3. Introduction of a national curriculum. The pre-reform compulsory schools were administrated by municipalities and the pre-reform curriculum varied between municipalities. 3.2 The Social Experiment The social experiment with the new comprehensive nine-year compulsory school started during an assessment period between 1949 and 1962, when the nal cur- 5 riculum was decided. The proposed new school system, as described above, was introduced in municipalities or parts of city communities, which in 1952 numbered 1,055 (including 18 city communities). The selection of municipalities was not random. However, the decision to select the areas was based on an attempt to choose areas that were representative for the entire country, both in terms of demographics as well as geographically. At rst the National Board of Education contacted the 4 The school reform and its development are described in Meghir and Palme (2003), Meghir and Palme (2005), and Holmlund (2007). For more detailed reference on the reform, see Marklund (1980) and Marklund (1981). 5 The ocial evaluation National School Board (1959) was mainly of administrative nature. Details on this evaluation are also described in Marklund (1981). 11 municipalities, or sometimes they themselves applied to participate. From this pool of applicants a "representative" sample of municipalities was chosen. Table 12 in the Appendix shows the take up rates of the reform between 1949 and 1962. In our analysis we consider cohorts born between 1945 and 1955. Figure 1 shows the number of observations in each one-year birth cohort and the proportion assigned to the reform. Figure 1: Proportion of individuals in sample assigned to the reform 4 The Data 4.1 Data Sources and Descriptive Statistics We use a sample originally obtained from Sweden's population census. To link individuals across generations we used the multi-generation register, provided by 6 Statistics Sweden. We are able to link and use three generations in our analysis: 6 Statistics Sweden(2003) Flergenerationregistret 2002. En beskrivning av innehåll och kvalitet. Statistics Sweden. Avdelning för Befolknings och Välfärdsstatistik. 12 the parent generation which is the generation directly aected by the reform and it consists of all individuals born in Sweden between 1945 and 1955, their parents the grandparent generation, the parent generation.7 655,801 females in the the children generation which are the children of This corresponds to 1,340,857 persons, 658,056 males and parent generation. date and parish of birth. generation, and the From the birth certicates we know gender, We restrict our sample of the children of the children generation, parent to those who have reached the age of criminal responsibility (age 15) in 2008, the last year for which we have crime records. This corresponds to 1,621,758 children, 833,564 sons and 788,194 daughters in the generation children that were born between 1959 and 1991. The reform assignment variable is obtained in two steps. First, we use the name of the church parish of birth in order to obtain the municipality code according to the 1953 Swedish municipality division. Second, based on the year and municipality of birth, we use an algorithm based on historical evidence on reform implementation in each municipality provided by Helena Holmlund and described in Holmlund (2007) to assign reform status to each individual in the sample. Information on the individual's highest education level was obtained and matched on to our sample from the Swedish National Education Register. For the ent generation grandpar- we used data from the 1970 census, which only provides information on individuals younger than age 60 in the year of the census, allowing us to obtain education information for 78.4 percent of mothers and for 65.8 percent of fathers of the parent generation. We analyze the eects of the reform separately for those individuals originating from the ned as those individuals of the low educated grandparent generation. grandparent generation 7 Even This is de- with the lowest pre-reform though we have information on biological and adoptive parents and children, we exclude all individuals who have been adopted, or who have adopted children themselves. 13 statutory level of compulsory schooling. Hence, we analyze the eects of the reform separately for the parent generation with low educated fathers, and for the children generation with low educated grandfathers, which amounts to roughly 63-65 percent 8 of the sample with available education information. Information on all convictions in entire Sweden covering the time period between 1981 and 2008 is provided by the Swedish National Council for Crime Prevention (Brå) and has been linked to individuals in our data set. One conviction/court trial often covers several crimes. We have detailed information on the number of crimes the person has been convicted for in each trial, the date of conviction, as well as the penalty for each crime. Table 1: Number of all convictions in Sweden 1981-2008, for cohorts born 1945-1955 and their children Ever convicted Ever convicted to prison Male Female Male Female Number 173,395 46,633 36,870 3,126 Percent 25.31 7.11 5.38 0.48 Number 68,860 16,673 14,525 1,255 Percent 23.96 6.11 5.05 0.46 Number 220,494 69,843 28,588 2,001 Percent 26.45 8.86 3.43 0.25 Parent generation Low educated fathers Children generation Low educated paternal grandfathers Number 64,976 20,183 7,361 482 Percent 24.68 8.12 2.80 0.19 Low educated maternal grandfathers Number 76,887 22,621 9,554 625 Percent 27.78 8.65 3.45 0.24 Notes: We only report the number of convicted children for low educated grandfathers, on both sides of the family. 8 Table 13 in the data appendix summarizes the number of available observations in each generation and subgroup. 14 Table 1 shows the number of convicted persons for the two generations, the 19451955 cohorts and their children, covered by our data on convictions. Over this time window, 25 percent of all males in the parent generation have been convicted at least once, and over 5 percent have been to prison. Only 7 percent of women have been convicted, and 0.5 percent have received a prison sentence. Importantly, the data on criminal convictions only cover the time range between 1981 to 2008, which means that the generations born between 1945 and 1955 will be between the ages of 26 to 63, whereas their children's convictions cover the ages of 15 to 49. The picture for the children generation looks very similar to the one of their parents, with slightly higher percentages of the population having been convicted, possibly attributable to the younger age window. Conviction rates for those with low educated fathers or grandfathers are presented in the lower panel of Table 1. This shows, perhaps surprisingly, that the conviction rates are only marginally higher for this group than for the whole population. Table 14 in the data appendix shows the crime-age distribution for the entire data set not only covering the cohorts of interest. The largest amount of convictions are for people between 15 and 24, followed by the age range 25 to 34, and further decreasing with age. This pattern of convictions by age is also shown in gures 1 and 2 in the appendix that show the average rate of convictions by age and by cohorts for the cohorts 1970-1989 using men in our children sample. The stated conviction rates for men of roughly 25 percent is a surprisingly high proportion of the population, which prompted us to look into this in greater detail. First, note that the type of crimes included in our data have to be severe enough to involve a trial and a conviction in court. This includes the more serious trac violations such as driving without a licence, driving under the inuence of alcohol or 15 drugs, and causing bodily harm, but does not include speeding or parking tickets. As such they do represent serious anti-social behavior. Unfortunately, we were not yet given the specic type of crime for which an individual in our data has been convicted for. 9 However, a good idea of the composition of crime can be obtained in Table 16 in the data appendix where we show a breakdown of type of crime convictions in 2009. In addition to the data on convictions we have data on all suspected crimes between 1991 and 2009. It includes a variable that gives a detailed code on the type 10 of suspected crime. Although this data overstates actual charges and crimes we use it to provide an idea of the distribution of trac crimes. Table 17 presents all categories that are related to trac violations and the number of oences between January 1991 and June 2009. The total number of suspected crimes during this time were 4,073,985 of which 16.9 percent were trac crimes. Again, all of these trac crime categories are severe violations. Additional support of such high conviction rates in Sweden is provided by previous studies that have shown similar conviction rates, see Hjalmarsson and Lindquist (2010a) and Hjalmarsson and Lindquist (2010b). 4.2 The Association of Education, Family and Criminal Behavior Table 2 shows the Linear Probability Model estimation results of whether an individual has ever been convicted (column (1)) and ever been convicted to a prison sentence (column (2)) on years of own schooling as well as on years of father's and 9 We are waiting to obtain a variable that indicates the type of crime from the Brå crime registry. coding of crime types in: Kodning av brott, Anvisningar och regler, Version 8.0, Reviderad 1. Juli 2010, brå brottsförebyggande rådet. 10 Detailed 16 mother's schooling for the children generation. The estimations for the sample of men born in Sweden between 1945 and 1955 are presented in the upper panel A, and the ones for the sons in the lower panel B. All regression models include both a full set of dummy variables for the birth cohorts and birth municipalities to control 11 for area and cohort xed eects. Column (1) in panel A shows that one year of own schooling for men in the parent generation is associated with a decrease of the probability of a conviction by 1.9 percentage points, which translates to a reduction in the conviction rate of 7.5 percent. One year of schooling for men decreases the probability of a prison sentence by 0.8 percentage points, or a reduction in the total share of men convicted to prison by 15 percent, see column (2). Panel B includes own schooling and both parents' education for sons, and hence estimates the eect of an increase of schooling of one parent controlling for assortative mating eects. One year of own schooling is associated with a 2.64 (0.59) percentage points decrease in the probability of sons' conviction (imprisonment), corresponding to a 9.96 (17.06) percent reduction in the share of convicted (imprisoned) sons. A years decrease in father's schooling is associated with an increase in the son's conviction rate (prison rate) of 0.568 (0.118) percentage points or equivalently a 2.15 (3.44) percent reduction in the share of convicted (imprisoned) sons. The equivalent gure for a decrease in mother's education by a year is a 2.261 (4.198) percent decrease in the share of convicted (imprisoned) sons. We present the relationship between crime and the levels of education in the data appendix (see Tables 18 and 19), revealing a steep decline in crime participation associated with higher levels of own and parental education. A similar decline is also recorded for incarceration rates. 11 When computing the standard errors we cluster by birth municipality, which is conservative and makes the standard errors comparable to the ones we estimate for the eect of the reform below. 17 Table 3 shows the results of a Probit model of the association between having a father who has been convicted (or to prison) and the father's age of birth on the probability of his son ever been convicted (or to prison). The probability of ever been convicted increases by over 15 percentage points if a son has a convicted father. This corresponds to a 63 percent increase of the total share of convicted sons. If he had a father who had been convicted to a prison sentence the probability of an own prison sentence increases by over 7 percentage points, which translates to a 134 percent increase in the share of convicted sons. Overall, there is a large and signicant negative correlation between own education and own criminal participation, both for the probability of a conviction and the probability of a conviction to prison. This holds for two subsequent generations. Furthermore, our evidence strongly supports the notion that higher education of both mothers and fathers is associated with a lower probability that their sons will be engaged in crime. Finally, we nd a strong and highly signicant positive intergenerational link between father's and son's criminal behavior. 18 Table 2: Linear probability model estimates of the association between own or father's education and the probability of being convicted and the probability of being convicted to a prison sentence. Men born between 1945-1955 and their sons. Panel A Dependent variables Men born 45-55 Probability conviction Probability prison p̄ = 0.2531 p̄ = 0.0538 -1.916*** -0.806*** (0.093) (0.057) -7.570 -14.981 y y 662,875 662,875 Years of schooling, own Corresponding percentage change Birth cohort/municipality dummies Observations Panel B Dependent variables Sons of men born 45-55 Probability conviction Probability prison p̄ = 0.2645 p̄ = 0.0343 -2.635*** -0.585*** Years of schooling, own Corresponding percentage change Years of schooling, father Corresponding percentage change Years of schooling, mother Corresponding percentage change Birth cohort/municipality dummies Observations (0.029) (0.013) -9.962 -17.055 -0.568*** -0.118*** (0.035) (0.014) -2.147 -3.440 -0.598*** -0.144*** (0.026) (0.009) -2.261 -4.198 y y 675,625 675,625 Notes: Results are scaled by 100. Level of observation in upper panel are men born in Sweden between 1945 and 1955. Level of observation in lower panel are sons of parents born between 1945 and 1955. Robust standard errors in parentheses, clustered by birth municipality. All regressions in the upper panel include a full set of birth cohort dummies and birth municipality dummies. In the lower panel a full set of municipality dummies and birth cohort dummies of the child was included. Signicance levels *** p<0.01, ** p<0.05, * p<0.1. 19 Table 3: Probit model marginal eects of the association between the probability of ever been convicted and the father having ever been convicted. Marginal eects are scaled by 100. Estimations for dierent education levels of grandfathers. Sons of men born in Sweden between 1945 and 1955. Panel A Dependent variables Father convicted/Father prison Percent change Age of father at birth of son Observations Sons of men born 45-55 Probability conviction Probability prison p̄ = 0.245 p̄ = 0.054 15.457*** 7.248*** (0.260) (0.183 ) 63.090 134.222 -1.759*** -0.323 *** (0.019) (0.007) 559,085 559,085 Panel B Sons of men born 45-55 with low educated father Dependent variables Probability conviction Probability prison p̄ = 0.245 p̄ = 0.021 14.755*** 7.116*** (0.245) (0.243) 60.224 338.857 -1.765*** -0.315*** Father convicted/Father prison Percent change Age of father at birth of son Observations (0.022) (0.007) 241,716 241,716 Notes: Level of observation in upper panel are sons of men born in Sweden between 1945 and 1955, and in the lower panel sons of men born between 1945 and 1955 who have a low educated father. Robust standard errors in parentheses, clustered by birth municipality of sons. All regressions include a full set of birth cohort dummies of fathers. Signicance levels *** p<0.01, ** p<0.05, * p<0.1. 20 5 Identication Strategy The main outcome variable we use is whether an individual was ever convicted during the observation window 1981-2008. We also consider whether someone has ever received a prison sentence. Finally, we consider the number of convictions. All the analysis is done for males only and we distinguish them by the education of the grandparent generation. We present two sets of estimates. The rst relate to the impact of the reform on the parent generation, i.e. the generation aected by the educational reform directly. The second relate to the impact of the reform on the children of the parent generation. Referring back to the discussion on how the reform aects crime, for the parent generation the estimated impact related to adult crime only because the youngest person in the sample is 26 when the crime records start. Hence, the eect is not attributable to the simply keeping the kids o the streets by getting them to attend school. For the child generation the impact is attributable to better parenting and increased investments in children; it is not due to dierences in the schooling system they are facing, which is the same for the treatment and comparison group children. If the reform had been randomised we could just compare crime rates across treatment and control municipalities at each point in time. However such randomisation did not take place and as a result we need further identifying assumptions so as to account for potential dierences across treatment and control municipalities. Assuming that the propensity to commit a crime can be written as ∗ = F (, t), yi,m,t where t denotes cohort and denotes unobserved heterogeneity. As in Athey and 21 Imbens (2006) and Altonji and Blank (1999) the key assumptions are that (i) the distribution of , while diering across municipalities, does not change from cohort to cohort and (ii) F is strictly monotonic in . These "dierence-in-dierences" assumptions are sucient to identify the impact of the reform, but do imply restrictions on the model for crime participation equation 2.3. In particular it requires that if there is any unobserved heterogeneity, whose distribution depends on the municipality, this has to be summarised by a single unobserved factor, say entering through human capital, and whose distribution is the same across cohorts; the propensity to commit a crime must be strictly monotonically related to this unobserved factor. In practice this assumes that increases in human capital always reduce crime and that this is never reversed by the reduction in the probability of being caught, which does not seem to be a very strong assumption. With a discrete outcome variable, such as ours we need additional assumptions if 12 we are to obtain a point estimate. We thus assume that heterogeneity and cohort eects are additively separable and we use both the linear probability model and the logit model as alternative distributional assumptions. Specically we assume ∗ yi,m,t = α + β1 Ri,m,t + γ10 ti + γ20 Mi + i,m,t , where ∗ yi,m,t is the latent crime "intensity" outcome observed for person municipality m and in birth cohort t. A conviction corresponds to is the reform indicator, which equals one if individual i i ∗ yi,m,t > 0. Ri,m,t ti is a vector of belongs to and vector of dummy variables indicating in which municipality individual 12 This born in belongs to a municipal- ity and cohort that has been assigned to the new school system; dummy variables indicating to which cohort individual i i Mi is a was born. is the case with all papers that employ dierence-in-dierences analysis of discrete outcomes (see Athey and Imbens (2006)). 22 Identication follows by assuming that E(i,m,t |Ri,m,t , ti , Mi ) = 0. The key assumptions are that the distribution of the unobservables are invariant over time and that the crime trends in the treated groups are the same as those in the comparison groups. The long time series of the data where we observe criminal activity later in life of cohorts who were not aected by the reform but went to school in municipalities that later implemented the reform and indeed also for several cohorts that went to the new school system allows us to present some corroborative evidence in support of the common trends assumption. Our rst approach is the linear probability model where ∗ Pr(yi,m,t > 0|Ri,m,t , ti , Mi ) = α + β1 Ri,m,t + γ1 ti + γ2 Mi . This is straightforward to estimate by GLS. However, the linear probability model has the well known disadvantage that it does not restrict the probabilities to lie on the [0,1] interval. The main reason we use it is for computational convenience, given the fact that there are about 1,000 municipality and 11 cohort xed eects. As an alternative, we also estimate a logit model by minimum distance: rst we group the data by municipality and cohort and estimate the within-cell conviction probability (Pmt ). We then use minimum distance to impose the restriction that this probability is generated from a logistic distribution with a linear index as in the latent equation above by tting the log-odds ratio as follows 23 log( Pmt ) = δ0 + δ1 Rm,t + δ20 ti + δ40 Mi . 1 − Pmt We drop all cells where the log odds ratio is not dened, which amounts to about 6% of cells. Implicitly the logit and the LP models deal with such cells as well as with the nonlinear form of the probabilities in a dierent way and hence it is interesting to see if we get dierent results with the two approaches. 6 Results 6.1 The First Generation Tables 4 through 7 show the estimates of the causal eect on criminal behavior of being assigned to the reform. Tables 4 and 5 show the results for the probability of having ever been convicted for the linear probability and logit model, respectively. Table 6 shows the results for the probability of having ever been convicted to a prison sentence using the linear probability model and, nally, Table 7 shows the eect of the reform on the intensity to engage in crime measures by the number of crimes an individual was convicted for. Throughout, we present results for the entire sample of men as well as separately for those coming from homes with low educated fathers. All specications include xed eects for birth municipality as well as birth cohorts and the standard errors are corrected for clustering within municipality of 13 birth, allowing for both spatial and serial correlation. To avid the computational diculties involved in estimating a logit model with 1,000 municipality xed eects and 11 cohorts we use a minimum distance procedure. 13 On clustering the standard errors see Moulton (1990). 24 We collapse the sample to 10,744 municipality-cohort cells by computing the logodds ratio within each cell. For 691 municipality-cohort cells the proportion of observed crime was zero and hence the log-odds ratio is not dened. For 108 cells we cannot assign the reform status, which leaves us with 9,949 municipality-cohort observations. We then regress the log-odds ratio on the municipality and cohort dummies as well as on the reform indicator using GLS. Each cell was weighted by p pc (1 − pc )Nc , where Nc is the cell size and pc is the within cell probability of a conviction. The corresponding marginal eects for dierent cohorts of the logit model are presented in Table 5. For the estimates when prison sentences are used as dependent variable this is not possible since the proportion of prison sentences is too small, which prevents us from computing the log-odds ratios. Column (1) in Tables 4 and 5shows a negative, but insignicant eect of the reform on the probability of having ever been convicted for the entire sample born between 1945 and 1955 and, in the lower panel of the tables, a similar result for those originating from homes with low educated fathers. However, these results may be diluted by the fact that the oldest birth cohorts in our data are quite old when we start to collect data on convictions and therefore may have ended their criminal careers. For example, the oldest cohort included, those born in 1945, are aged 36 when we start to record their criminal behavior. This is an age by which a large proportion have already stopped committing crimes. To account for this possibility, we also present results where we have restricted the sample to younger birth cohorts. Columns (2) through (6) in Tables 4 and 5 show results for the impact of the reform on dierent cohort groups. Although, we clearly loose some precision as we reduce the number of cohorts considered, the results show a clear pattern for both the entire sample and those originating from 25 homes with low educated fathers: for younger cohorts, which are observed at the age range when more crimes are committed, the reform seems to have increasingly strong eects. The estimate for the youngest cohort, born in 1954 or 1955, is highly signicant suggesting a 1.3 percentage points decrease in crime; this corresponds to a 5 percent, decrease in the probability of ever been convicted as a result of being assigned to the post reform school system. The eect is somewhat larger in the subsample of men born in homes with low educated fathers. These results are consistent with ndings by Lochner and Moretti (2004) and Machin, Marie, and Vuji¢ (2010) for the US and the UK respectively, who show that increases in compulsory schooling led to reductions in crime. To analyze if the education reform also decreased the amount of crimes individuals were convicted for we use an alternative outcome variable that measures the total number of crimes an individual has been convicted for between 1981 and 2008. The average number of convicted crimes for men in the rst generation ranges from 1.2 to 1.8, depending on cohort and father's education level. This outcome variable is characterized by a large mass point at zero and overdispersion with a variance roughly ten times larger than the mean. To estimate the model on this outcome variable we use a negative binomial model. Table 7 presents the marginal eects from these estimates. The estimates are negative and signicant at the 10 percent level for those born between 1950 and 1955 in homes with low educated fathers. The eect seems to be somewhat stronger for those born between 1952-1955 and 1953-1955, respectively. These results are in line with the ndings on the probability of having ever been convicted. 26 Table 4: Linear Probability Model estimates of the eects of the reform on the probability of ever been convicted. Marginal eects are scaled by 100. Estimations by birth cohorts for the entire sample and for men with low educated fathers separately. Men born in Sweden between 1945 and 1955. Panel A Cohorts Men born 45-55 45-55 50-55 51-55 52-55 53-55 54-55 Probability conviction 0.2531 0.2683 0.2714 0.2746 0.2783 0.2837 Reform -0.645 -0.456 -0.532* -1.028*** -1.076** -1.329*** (0.405) (0.305) (0.318) (0.396) (0.490) (0.479) -2.548 -1.700 -1.960 -3.744 -3.866 -4.685 622,583 319,093 263,592 210,399 157,155 103,761 Percentage change Share convicted Observations Panel B Cohorts Men born 45-55 with low educated father 45-55 50-55 51-55 52-55 53-55 54-55 Probability conviction 0.2396 0.2517 0.2552 0.2588 0.2631 0.2693 Reform -0.263 -0.494 -0.579 -1.119* -1.103 -2.094** (0.304) (0.434) (0.486) (0.661) (0.912) (0.970) -1.098 -1.963 -2.269 -4.324 -4.192 -7.776 264,679 150,620 125,952 101,266 76,207 50,222 Percentage change Share convicted Observations Notes: The table presents results p from a linear probability (LP) estimation with weighted least squares. The used weights are: x0 b(1 − x0 b), which are obtained from a rst stage OLS estimation. The level of observation are men born between between 1945-1955, or men born in cohorts indicated at the top of each column; Reform is a dummy variable indicating the reform status of the index person. All regressions include a full set of birth municipality dummies and birth cohort dummies. Robust Standard errors in parentheses, clustered by municipality of birth. *** p<0.01, ** p<0.05, * p<0.1. 6.2 The Child Generation Tables 8 and 9 report the results of the dierence-in-dierences estimation of the eects of the school reform on criminal participation of the child generation for the probability of ever been convicted and prison convictions, respectively. Again, for the probability of being convicted, we additionally estimate a logit model based on cohort-municipality cells. We estimate two specications. In the rst one, we estimate the eects of a father who has been subjected to the new school system on 27 Table 5: Logit model marginal eects of the reform on the probability of ever been convicted. Marginal eects are scaled by 100. Estimations by birth cohorts for the entire sample and for men with low educated fathers separately. Men born in Sweden between 1945 and 1955. Panel A Men born 45-55 Cohorts 45-55 50-55 51-55 52-55 53-55 54-55 Probability conviction 0.253 0.268 0.271 0.275 0.278 0.284 Reform -0.023 -0.284 -0.434 -1.274* -0.999 -1.395 (0.281) (0.422) (0.500) (0.654) (0.911) (1.361) -0.091 -1.057 -1.598 -4.639 -3.591 -4.916 622,583 319,093 263,592 210,399 157,155 103,761 Percentage change Share convicted Observations Panel B Men born 45-55 with low educated father Cohorts 45-55 50-55 51-55 52-55 53-55 54-55 Probability conviction 0.234 0.252 0.255 0.259 0.263 0.269 Reform -0.022 -0.272 -0.417 -1.227* -0.965 -1.351 (0.271) (0.405) (0.480) (0.630) (0.879) (1.318) -0.093 -1.081 -1.633 -4.741 -3.666 -5.015 264,679 150,620 125,952 101,266 76,207 50,222 Percentage change Share convicted Observations Notes: We only report marginal eects of the logit model; Reform is a dummy variable indicating the reform status of the index person. All regressions include a full set of birth municipality dummies and birth cohort dummies. The level of observation are men born between 1945-1955, or men born in cohorts indicated at the top of each column. Robust Standard errors in parentheses, clustered by municipality of birth of the individual. *** p<0.01, ** p<0.05, * p<0.1. child criminal behavior and in the second one the corresponding eects of a mother going through the post reform school system. We distinguish between four samples. The rst is the sample of individuals whose father was born between 1945 and 1955 for which we estimate the eect of their father's school reform assignment on their probability of a conviction. The second sample are individuals whose mother was born between 1945 and 1955, for which we estimate the eect of their mother's school reform assignment on their son's criminal behavior. The other two samples are individuals originating from the 28 low educated Table 6: Linear Probability Model estimates of the eects of the reform on the probability of ever been convicted to a prison sentence. Marginal eects are scaled by 100. Estimations by birth cohorts for the entire sample and for men with low educated fathers separately. Men born in Sweden between 1945 and 1955. Sample Probability prison Reform Corresponding percentage Men born 45-55 Men born 45-55 with low educated fathers p̄ = 0.0538 p̄ = 0.0505 -0.149 -0.049 (0.160) (0.170) -2.770 -0.974 622,583 264,679 change share prison Observations Notes: The table presents p results from a linear probability (LP) estimation with weighted least squares. The used weights are: x0 b(1 − x0 b), which are obtained from a rst stage OLS estimation. The level of observation are men born between 1945-1955. 5.38% of all men in the considered sample have been convicted to prison at least once. 5.05% of the considered sample who have a low educated father have been to prison. Robust Standard errors in parentheses, clustered by municipality of birth. *** p<0.01, ** p<0.05, * p<0.1. Reform is a dummy indicating the reform status of the index person. All regressions include a full set of birth municipality dummies and birth cohort dummies. grandparent generation. More specically, the third sample is dened like the rst one, but restricted to those individuals whose grandfather on the father's side is low educated. Finally, the fourth sample is restricted to those individuals within the second sample whose grandfathers on their mother's side are low educated. When we collapse the data set by cohort-municipality level as a rst step to estimating the logit model, we do it by father's cohort-municipality level which gives us 10,607 cells for the father's sample, and 10,247 for the sample with low educated paternal grandfathers. For the specication with mother's reform assignment we collapse the data by mother's cohort-municipality level which leads to 10,647 for the entire sample and 10,324 for the low educated maternal grandfathers sample. Table 8 reports the marginal eects of the linear probability model for a conviction (denoted by LP), the log-odds ratio estimate for the logit model, and the corresponding marginal eect of the logit model, both for the sample of all sons, and for sons whose grandfather was low educated. It can be seen that the reform 29 Table 7: Marginal eects of negative binomial estimation of the eects of the reform on the number of crimes men have been convicted for between 1981-2008. Estimations by birth cohorts for the entire sample and for men with low educated fathers separately. Men born in Sweden between 1945 and 1955. Sample Men born 45-55 Cohorts 45-55 50-55 51-55 52-55 53-55 54-55 ȳ 1.309 1.578 1.646 1.696 1.748 1.828 Reform 0.027 -0.071 -0.038 -0.122 -0.078 -0.121 (0.035) (0.062) (0.067) (0.086) (0.083) (0.092) 622,583 319,093 263,592 210,399 157,155 103,761 Observations Sample Men born 45-55 with low educated fathers Cohorts 45-55 50-55 51-55 52-55 53-55 54-55 ȳ 1.253 1.473 1.539 1.587 1.647 1.717 Reform -0.030 -0.137* -0.117 -0.250** -0.236* -0.177 (0.047) (0.081) (0.096) (0.124) (0.128) (0.168) 264,679 150,620 125,952 101,266 76,207 50,222 Observations Notes: Margnial eects of the negative binomial estimation are reported. Reform is a dummy indicating the reform status of the individual. Robust Standard errors in parentheses, clustered by individual's municipality of birth. *** p<0.01, ** p<0.05, * p<0.1. All regressions include a full set of birth municipality dummies and birth cohort dummies of the index person. The estimation method is a negative binomial model to relax the Poisson model assumption of equal conditional mean and conditional variance of the outcome variable, but maintaining an exponential mean. signicantly reduces the probability of having ever been convicted for the sons of fathers who went through the reform by 0.6 percentage points, see column (LP). Since the average share of convicted individuals in this cohort was about 26.5 percent, reduction in criminality was approximately 2.5 percent. The log-odds-ratio estimates from the logit model translate into a marginal effect of a 0.646 percentage points decrease in the probability of a conviction, which is very similar to the marginal eect obtained from the linear probability model. Hence, the logit model suggests similar to the linear probability model, that father's reform assignment signicantly reduces the total share of convicted men by about 2.5 percent. 30 Consider now the sample of sons with a low educated paternal grandfather. The impact of their father being assigned to the reform is a reduction of the probability of conviction by 1.02 percentage points. This translates into a 4.13 percent decrease in the share of convicted sons of low educated grandfathers (24.7 percent). The marginal eect of the logit model is of the same magnitude (-0.97 percentage points, and -3.94 percent decrease in total share of convicted). There are no signicant eects of reform assignment of mothers on the probability of a conviction of their sons neither for the entire sample of sons, nor for those with a low educated maternal grandfather. Also, as can be seen in Table 9, the results of the linear probability model for the eects of the reform status of father and mother on the probability of a prison sentence of sons show no signicant eects. Table 10 shows the negative binomial estimates of the marginal eects for the intensive margin on the number of crimes convicted for between 1981 and 2008 for the children generation. The top panel A shows the results based on father's reform status for all children and those with low educated grandfathers, respectively. The bottom panel B shows the corresponding results based on mother's reform status. The estimate based on father's reform status shows a signicant negative eect of the reform based on father's reform status, which is in line with our ndings on the probability of being convicted. We repeated the analysis for the children generation using the suspected crime data mentioned above. Descriptive tables on suspected crime rates for the relevant sample of children is presented in the appendix in table 15. Interestingly, the rates of ever been suspected for a crime including and excluding all trac crimes are very similar. We nd a negative but not signicant relationship between father's reform assignment and sons probability of ever been suspected for a crime. 14 The results are provided by the authors upon request. 31 14 32 595,138 595,138 (0.014) (0.249) -0.600 0.012 Logit 595,138 0.851 (0.277) 0.225 Logit Marginal eect 255,075 -0.147 (0.331) -0.041 LP 255,075 (0.018) 0.021 Logit 255,075 1.509 (0.363) 0.419 Logit Marginal eect p̄ = 0.278 243,082 -3.938 (0.366) -0.972*** Sons of women born 45-55 with low educated fathers 243,082 243,082 (0.020) (0.361) -4.129 -0.052*** -1.02*** Logit Marginal eect p̄ = 0.265 563,754 -2.443 (0.296) -0.646*** Logit Sons of women born 45-55 -0.159 LP 563,754 563,754 (0.015) (0.219) -2.456 -0.033** LP p = 0.247 Logit Marginal eect p = 0.265 Logit -0.650*** LP Sons of men born 45-55 with low educated fathers Sons of men born 45-55 Notes: The top panel shows the estimated eects of father's reform status Reform_father on sons crime, and the bottom panel shows the estimated eects of mother's reform status Reform_mother on son's crime. 26.45% of men in the considered sample have been convicted at least once. Of those men who have a low educated grandfather on their father's side 24.68% have been convicted. Of those men who have a low educatedp grandfather on their mother's side 27.78% have been convicted. Columns (1) and (4) present linear probability model (LP) estimation with weighted least squares. The used weights are: x0 b(1 − x0 b), which are obtained from a rst stage OLS estimation. The level of observation are sons born before 1994, whose parents were born between 1945-1955. Column (2) and (5) show the results of a logit estimation; we report the coecient of the log odds ratio coecient. The marginal eect for the logit model are presented in columns (3) and (6). We reduced the sample on father's or mother's birth municipality and birth cohort level to estimate the logit model. Reducing the sample to the following parent's cohort-birth municipality cells: for sons with a father born 1945-1955 to 10,607 cells, for sons whose father has a low educated father to 10,247 cells, for sons whose mother was born 1945-1955 to 10,647 cells, and whose mother had a low educated father to 10,324 cells. Robust Standard errors in parentheses, clustered by municipality of birth. *** p<0.01, ** p<0.05, * p<0.1. All regressions include a full set of birth municipality dummies and birth cohort dummies of father's or mother's. Observations Change share convicted Corresponding percentage Reform_mother Specication Probability conviction Sample Panel B: Mother's reform assignment Observations Change share convicted Corresponding percentage Reform_father Specication Probability conviction Sample Panel A: Father's reform assignment with low educated grandfathers separately. Sons of men and women born in Sweden between 1945 and 1955. the probability of their sons having ever been convicted. Marginal eects are scaled by 100. Estimations for the entire sample and for sons Table 8: Linear Probability Model estimates and Logit Model estimates of the marginal eects of father's and mother's reform assignment on Table 9: Linear Probability Model estimates of the eects of father's and mother's reform assignment on the probability of their sons having ever been convicted to a prison sentence. Marginal eects are scaled by 100. Estimations for the entire sample and for sons with low educated grandfathers separately. Sons of men and women born in Sweden between 1945 and 1955. Panel A: Father's reform assignment Sample Probability prison All Sons Sons with low educated grandfather p = 0.034 p = 0.028 0.015 0.005 (0.093) (0.123) 0.448 0.169 563,754 243,082 All Sons Sons with low educated grandfather p = 0.034 p = 0.028 Reform_father Corresponding percentage Change share prison Observations Panel B: Mother's reform assignment Sample Probability prison Reform_mother 0.038 0.004 (0.093) (0.130) 1.093 0.108 595,138 255,075 Corresponding percentage change share prison Observations Notes: The reported estimates are based on a weighted least squares estimation. 3.43% of all men in the sample have been to prison. 2.80% of men who have a low educated grandfather on their father's side have been convicted to prison. 3.45% of men who have a low educated grandfather on their mother's side have been convicted to prison. All regressions include either a full set of birth municipality dummies of father or mother, and birth cohort dummies of father or mother. Robust Standard errors in parentheses, clustered by municipality of birth of mother or father. *** p<0.01, ** p<0.05, * p<0.1. A possible mechanism for the inverse eect of the reform on the criminal behavior of the child generation is through fertility and family formation. An implication of both the well known quality-quantity model and the time allocation model is that increased educational attainments and earnings prospects will decrease the demand for having children. If there is a trade o between the number of children and parental investments in each child, it may make up an inverse relation between 33 Table 10: Marginal Eects of negative binomial estimation of father's or mother's reform assignment on the number of crimes sons are convicted of between 1981-2008. Estimations for the entire sample and for sons with low educated grandfathers separately. Sons of men and women born in Sweden between 1945 and 1955. Panel A: Father's reform assignment Sample All Sons Sons with low educated grandfather 1.245 1.210 -0.065** -0.040 (0.031) (0.050) 563,754 243,082 All Sons Sons with low educated grandfather Average number convicted crimes 1.538 1.442 Reform_mother 0.010 0.069 (0.046) (0.055) 595,138 255,075 Average number convicted crimes Reform_father Observations Panel B: Mother's reform assignment Sample Observations Notes: Marginal eects are reported, ȳ is the average number of convicted crimes in the sample. Experiment_father or Experiment_mother is a dummy indicating the reform status of the father or mother of the son. Robust Standard errors in parentheses, clustered by municipality of birth of the son's father (Panel A) or mother (Panel B). *** p<0.01, ** p<0.05, * p<0.1. All regressions include a full set of birth municipality dummies and birth cohort dummies of father (Panel A) or mother (Panel B). The estimation method is a negative binomial model to relax the Poisson model assumption of equal conditional mean and conditional variance of the outcome variable, but maintaining an exponential mean. being assigned to the reform and the inverse relation to child criminality. To further investigate this possible mechanism we estimate the eect of the reform on fertility behavior. We use four dierent outcomes: the total number of children, the probability of having children, the age at birth of rst child and, nally, the probability of having children while still a teenager. Since we only nd signicant eects of the reform on sons to fathers aected by the reform, we restrict the sample to men. 15 As can be seen in the results of these estimates shown in table 11, the only signicant eect is on the probability of having a child as a teenager. 15 The results for women are qualitatively the same, the only dierence is that women who were assigned to the reform are less likely to ever have a child on a 10 percent signicance level. 34 However, since only 1.3 percent of the population have children as teenagers and, more importantly, the reform only decreased this probability by between 0.2 and 0.3 percent, we conclude that we cannot attribute the eect of the reform to changes in fertility behavior. The results on all other fertility outcome variables shown are estimated to be zero. Table 11: Estimation of the eects of the reform on the probability of ever having a child, the number of children, the age at birth of rst child and the probability of teenage paternity. Marginal eects are scaled by 100. Estimations by birth cohorts for the entire sample and for men with low educated fathers separately. Men born in Sweden between 1945 and 1955. Dependent variable Specication ever child number children age birth rst child teenage LP Poisson Neg binomial LP 0.813 1.896 Sample: Average dep var Reform Observations Men born 45-55 Reform Observations 0.017 -0.093 -0.004 0.106 -0.263** (0.185) (0.007) (0.075) (0.106) 622,583 622,583 505,679 622,583 Sample: Average dep var 27.054 Men born 45-55 with low educated fathers 0.822 1.912 26.524 0.019 -0.096 0.001 0.064 -0.210** (0.273) (0.010) (0.048) (0.100) 264,679 264,679 217,517 264,679 Notes: Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1. Stan- dard errors are clustered by municipality level. All estimations include a set of municipality dummies and cohort dummies. The level of observation are men born between 1945 and 1955, all men born between 1945-1955. Experiment is a indicator variable for reform assignment. 6.3 The common trends assumption As is well known, an identifying assumption underlying the dierences-in-dierences estimator is that any trend in the outcome variable is common in the treatment and comparison groups over the period of comparison. 35 This assumption is untestable because it relates to the counterfactual change in the treatment group. However, an indication can be obtained by testing whether the trends are common in the two sets of groups before the reform and indeed after the reform as well. In our sample we have 12 groups of municipalities indexed by which cohort was rst assigned to the reform. To implement the common trends test we used only the municipalities that rst implemented the reform for the 1947 cohort onwards (i.e. 10 of the 12 set of municipalities). We compared the trend of criminal behavior of individuals across these municipalities for all cohorts that were not aected by the reform. The pooled regression of these groups is where y = α + βt + γ 0 m + δ 0 t ∗ m + ν , m is a set of dummies indicating the group to which the municipality belongs based on the cohort for which it rst implemented the reform; that represents the cohorts 1945-1955. A joint test of F statistic of F (9;7,090)=1.15 with P =0.323, t is a linear trend δ1 = δ2 = ... = δ10 = 0 gives a with 7,090 cohort-municipality cells before treatment. This implies the hypothesis of common trends in crime for the pre-treatment cohorts for all groups of municipalities cannot be rejected. For post-treatment trends in crime we only compare crime between the municipalities that implemented the reform for cohorts born before 1954. We compare the criminal behavior of individuals across these municipalities for the cohorts aected by the reform. A joint test of equality of the coecients on the interaction term of the above pooled model yields F(9, 4,808) with P=0.1303, where 4,808 is the number of cohort-municipality cells that are treated. This means that the hypothesis of a common trend in criminal behavior for the treated cohorts is the same across the groups of municipalities that implemented the reform for dierent cohorts cannot be rejected. Both these tests are strong evidence in favor of the key identifying assumption for our dierence-in-dierences approach to the problem. 36 7 Conclusion We nd a very strong relation between own educational attainments and criminal behavior in our data: every year of additional education is associated with a decrease of about 7.5 percent in the probability of being convicted for a crime and a 15 percent decrease in the probability of a prison sentence. We also nd a strong relation between parental education and crime. On average, one year of additional education of the father is associated with an about 2.1 percent decrease in the probability of being convicted and an about 3.5 percent decrease in the probability of a prison sentence. To what extent do these strong correlations give guidance to the likely eects of a policy intervention designed to improve educational attainment of the least skilled? The staggered implementation of a major Swedish education reform in the 1950s and early 60s, which among other things increased compulsory schooling, provides an ideal opportunity to answer this question. Using administrative data that compares children of the same cohorts, but educated under dierent systems, we nd strong negative and signicant eects of the reform, particularly when we focus on cohorts whose criminal activity is recorded at a younger age. Thus, for the youngest cohorts, born between 1954 and 1955, the point estimate suggests a 1.3 percentage points, corresponding to 5 percent, decrease in the probability of being convicted from being assigned to the post reform school system. In the group from homes with low educated fathers the eect seems to be somewhat larger, which is consistent with a larger eect on both educational attainments and labor earnings in this group found in Meghir and Palme (2005). The striking result of this paper however is the eect of the reform on the children of those originally aected: there is a signicant eect of paternal assignment to the reform on the probability of being convicted corresponding to an average reduction of 37 about 2.5 percent. However, we found no signicant eect from mother's assignment to the reform. The intergenerational eect of education on criminal behavior can operate through several dierent channels. The result that the reform had a stronger eect on criminal behavior in the child generation - together with the result obtained in Holmlund, Lindhal, and Plug (forthcoming) that the reform had a weak, although signicant, eect on the educational attainments in the next generation - suggests that eect did not operate solely through increased education in the child generation. Another conceivable channel is through changes in fertility induced by the reform. Our data allowed us to estimate reform eects on four dierent outcomes related to fertility behavior. Since our results suggest that the eect operated through father's reform assignment, we restricted the sample to men. The only signicant eect was a decreased probability of having a child as teenager. However, the eect was far too small to serve as a major explanation of the decreased criminality in the next generation. Given the data at our disposal, there are several channels that we are not able to distinguish between. The most obvious one is that improved earnings possibilities provide the parents with improved economic resources to invest in the upbringing of their children. However, since crime is also a manifestation of a basic lack of social skills or other emotional decits it is also conceivable that the eect of parental education works through other channels than solely economic resources. A possible interpretation of the fact that we nd signicant eects of the father's rather than the mother's reform assignment on the son's criminal behavior is consistent with the notion that these alternative explanations may be important, in view of the fact that female labor force participation is very high in Sweden. 38 References and Altonji, J. G., R. M. Blank (1999): Chapter 48 Race and gender in the labor market, vol. 3, Part 3 of Handbook of Labor Economics, pp. 3143 3259. Elsevier. Athey, S., and G. W. Imbens (2006): Identication and Inference in Nonlinear Dierence-In-Dierences Models, Becker, G. (1981): Econometrica, 74(2), 431497. A treatise on the family. Harvard University Press, New York. Björklund, A., M. Lindahl, and E. Plug (2006): The Origins of Intergenera- tional Associations: Lessons from Swedish Adoption Data, Quarterly Journal of Economics, 121(3), 9991028. Carneiro, P., C. Meghir, and M. Parey (2007): Maternal education, home environments and the development of children and adolescents, IZA Discussion paper No. 3072. Cunha, F., and J. Heckman (2007): The Technology of Skill Formation, The American Economic Review, 97(2), pp. 3147. Freeman, R. B. (1996): Why Do So Many Young American Men Commit Crimes and What Might We Do About It?, The Journal of Economic Perspectives, 10(1), pp. 2542. (1999): The Economics of Crime, vol. 3, Part 3, Chapter 52 of Handbook of Labor Economics. Elsevier. Gallipoli, G., and G. Fella (2006): Education and Crime over the Life Cycle, Working Paper. 39 Gould, E. D., B. A. Weinberg, and D. B. Mustard (2002): Crime Rates and Local Labor Market Opportunities in the United States: 1979-1997, Review of Economics and Statistics, 84(1), 4561. Grogger, J. (1998): Market Wages and Youth Crime, Journal of Labor Eco- nomics, 16(4), pp. 756791. Hjalmarsson, R., and M. J. Lindquist (2010a): Exploring the Intergenerational Nature of Crime, Like Godfather, Like Son: Working Paper. (2010b): The Origins of Intergenerational Associations in Crime: Lessons from Swedish Adoption Data, working paper. Holmlund, H. (2007): A Researcher's Guide to the Swedish Compulsory School Reform, Swedish Institute for Social Research, Working Paper 9/2007. Holmlund, H., M. Lindhal, and E. Plug (forthcoming): The Causal Eect of Parent's Schooling on Children's Schooling: A Comparison of Estimation Methods, Journal of Economic Literature. Lochner, L. (2004): Education, Work, and Crime: A Human Capital Approach, International Economic Review, 45(3), pp. 811843. Lochner, L., and E. Moretti (2004): The Eect of Education on Crime: Evi- dence from Prison Inmates, Arrests, and Self-Reports, The American Economic Review, 94(1), pp. 155189. Machin, S., O. Marie, Education, Machin, S., and S. Vuji¢ (2010): The Crime Reducing Eects of CEP Discussion Paper No 979. and C. Meghir (2004): Crime and Economic Incentives, nal of Human Resources, 39(4), pp. 958979. 40 The Jour- Marklund, S. (1980): Från reform till reform: Skolsverige 1950-1975, Del 1, 1950 Skolöverstylrelsen och Liber Utbildningsförlaget. års reformbeslut, (1981): Från reform till reform: Skolsverige 1950-1975, Del 2, Försöksverksamheten, Stockholm: Utbildningsförlaget. Meghir, C., and M. Palme (2003): Ability, Parental Background and Educa- tion Policy: Empirical Evidence from a Social Experiment, IFS working paper, (WP03/05). (2005): Educational Reform, Ability, and Family Background, The Amer- ican Economic Review, 95(1), pp. 414424. Moulton, B. R. (1990): An Illustration of a Pitfall in Estimating the Eects of Aggregate Variables on Micro Units, The Review of Economics and Statistics, 72(2), pp. 334338. National School Board, R. (1959): Försöksverksamhet med nioårig skolplikt, Stockholm, Report 52. Sabates, R., Crime, and L. Feinstein (2008a): Educational Attainment and Juvenile British Journal of Criminology. (2008b): Eects of Government Initiatives on Youth Crime, Oxford Eco- nomic Papers, 60(3), pp. 462483. Schweinhart, Belfield, L., and J. J. Montie, M. Nores (2005): Z.-p. Xiang, W. S. Barnett, C. R. Lifetime eects: The High/Scope Perry Preschool Study Through Age 40. Ypsilanti, MI: High/Scope Press. Williams, J., and R. C. Sickles (2002): 41 An Analysis of the Crime as Work Model: Evidence from the 1958 Philadelphia Birth Cohort Study, of Human Resources, 37(3), pp. 479509. 42 The Journal 8 APPENDIX 8.1 Reform Appendix Table 12: Quantitative development of the comprehensive school experiment 1949 to 1962. Year Municipalities Cumulative Number Percentage share Number of Number of classes students 1949/50 14 1.3 172 2 483 1950/51 20 1.9 379 7 529 1951/52 25 2.4 682 14 635 1952/53 30 2.9 1 009 22 725 1953/54 37 3.5 1 525 35 784 1954/55 46 4.4 2 516 61 498 1955/56 59 5.6 3 394 84 941 1956/57 71 6.7 4 393 109 694 1957/58 96 9.1 5 702 143 370 1958/59 142 13.5 8 036 196 343 1959/60 217 20.6 11 191 266 042 1960/61 295 28.0 14 283 333 094 1961/62 415 39.4 18 665 436 595 Note: The 1952 division of municipalities (total: 1 052). Source: Marklund 43 8.2 Data appendix Table 13: The Sample Number observations All Male Female 1,340,857 685,056 655,801 Parent generation: Cohorts 1945-1955 Father's education available 881,742 452,433 429,309 Of which low educated father 560,273 287,396 272,877 63.54 63.52 63.56 1,621,758 833,564 788,194 Paternal grandfather's education available 802,451 412,619 389,832 Of which with low educated grandfather 511,980 263,319 248,661 63.80 63.82 63.79 Maternal grandfather's education available 836,632 430,357 406,275 Of which with low educated grandfather 538,228 276,779 261,449 64.33 64.31 64.35 Percent Children generation: All children of parent generation Percent Percent Notes: We only present the number of observations that are available on father's and grandfather's education level, because we will condition on father's or grandfather's education level in the analysis. We only have information on the highest level of education for those individuals that are not older than 60 years in the year of the 1970 census. We report the number of individuals in each sample, the number of individuals for whome we have information on the highest level of education on their fathers or grandfathers and the share of those for which we have this information with the lowest education level. For the children generation with low educated grandfathers on their father's side of the family, we consider those children whose father was born between 1945 and 1955. For the children generation with low educated grandfathers on the mother's side of the family we consider those whose mother was born between 1945 and 1955. 44 Table 14: Number of all convictions in Sweden between 1981-2008 Number of convicted persons Number of persons convicted for prison All Male Female 1,249,569 966,790 282,779 366,639 344,919 21,720 Number of convictions in total 3,014,811 2,534,337 480,474 Number of prison sentences in total 1,204,711 1,115,428 89,283 age 15 -24 1,128,125 950,413 177,712 age 25-34 710,177 605,445 104,732 age 35-44 577,693 483,821 93,872 age 45-54 355,396 296,971 58,425 age 55-64 161,367 133,788 27,579 age 65-80 76,296 59,138 17,158 > 5,757 4,761 996 Convictions by age group age 80 78% to 85% of the convictions are males. Table 15: Data on all suspected crimes in Sweden between 1991-2009. Sons of men or women born 1945-1955. Number of persons suspected for a crime 1991-2009 All crimes Excluding trac Excluding some trac 129,683 117,279 124,487 20.95 18.94 20.11 Sample: Sons of men born 45-55 Percent of sample Sample: Sons of men born 45-55 with low educated father Percent of sample 54,542 48,888 52,222 20.71 18.57 19.83 133,953 120,748 129,217 20.50 18.48 19.78 Sample: Sons of women born 45-55 Percent of sample Sample: Sons of women born 45-55 with low educated father Percent of sample 55,210 49,294 129,217 19.95 17.81 19.78 Notes: The category excluding all trac crimes excludes the crime categories listed in table 17. The category excluding some trac crimes excludes the categories "Driving without a license", "Allowed driving without license" and "Override provision". 45 Figure 2: Life cycle convicted crimes Figure 3: Life cycle convicted crimes 46 Table 16: Persons found guilty of criminal oences, by principal oence Number of Crimes against penal code Of which Crimes against life and health Share, % convictions, 2009 59,542 42.1 9,744 6.9 150 0.1 Of which Murder and man-slaughter Assault, gross assault 9,268 Sexual oences 1,090 6.5 0.8 Of which: Rape Theft, robbery, other oences of stealing 256 29,393 0.2 20.8 Of which: Theft, gross theft Petty theft Robbery, gross robbery Vehicle theft Fraud and other dishonesty Crimes inicting damage Violent threat to public servant Other 9,233 6.5 17,953 12.7 1,049 0.7 824 0.6 3,175 3,316 2,544 10,280 Crimes to other penal legislation Crimes against the Road trac oences act 2.2 2.3 1.8 7.3 82,035 57.9 13,253 9.4 47,020 33.2 Of which Drunken driving, gross drunken driving Crimes against the Narcotics drugs act Crimes against the Act on smuggling Other All crimes 18,525 2,076 14,414 141,577 13.1 1.5 10.2 100 Persons found guilty of criminal oences, by principal oence, 2009. Source: Kriminalstatistik, Rättsstatistisk årsbok, Statistisk årsbok, Statistiska Meddelanden (R 11 SM). 47 48 0.01 0.05 0.05 0.06 0.17 0.2 0.57 1.31 1.45 2.81 7.25 24.43 61.64 Notes: Criminal oences related to trac violations persons were suspected for. Register data on all crimes individuals have been suspected for between January 1991 and June 2009 in Sweden. 83 1,201 Involuntary manslaughter in connection with accident Other crimes against maritime law, imprisonment in the range of penalties 1,393 Causing bodily injury or disease related to trac accident 314 3,896 Cause of danger for another in connection with trac or in trac Maritime Act: Drunkenness 9,018 Gross negligence in trac 321 9,941 Allowed driving without a license Other trac oense, imprisonment in range of penalties 19,346 Hit and run 414 49,828 Driving or aggravated driving under the inuence of drugs alone Override provision 167,958 423,809 4,073,985 687,522 (16.876%) Number of cases Percentage of trac crimes or under the inuence of both alcohol and drugs Drink-driving or aggravated drink-driving under the inuence of alcohol alone Driving or aggravated driving without a license Total number of crimes Total number of trac crimes Description of trac violation Table 17: Trac crimes persons were suspected for, January 1991 - June 2009. Table 18: Linear probability estimates of the association of own education levels and criminal behavior. Men born between 1945-1955. Dependent variables Probability conviction Probability prison p̄ = 0.2531 p̄ = 0.0538 Education Levels Vocational Upper secondary Upper secondary + College/University PhD Observations ≥ 1 year -0.161 0.031 (1.190) (0.273) -7.471*** -3.928*** (1.028) (0.287) -10.549*** -5.113*** (0.913) (0.288) -13.782*** -5.929*** (0.923) (0.395) -19.759*** -7.183*** (0.713) (0.545) 684,625 684,625 Notes: Results are scaled by 100. Level of observation are men born between 1945 and 1955 in Sweden. Robust standard errors in parentheses, clustered by birth municipality. Signicance levels *** p<0.01, ** p<0.05, * p<0.1. All regressions include a full set of birth cohort dummies and birth municipality dummies. 49 Table 19: Linear probability model estimates of the association of parental education levels and own criminal behavior. Sons of parents born 1945-1955. Dependent variables Probability conviction Probability prison p̄ = 0.2645 p̄ = 0.0343 -2.075*** -0.694*** (0.214) (0.088) -8.083*** -2.342*** (0.339) (0.142) -9.719*** -2.457*** Panel A: Education levels father Vocational Upper secondary Upper secondary + ≥ 1 year College/University PhD Observations (0.388) (0.135) -12.535*** -2.900*** (0.463) (0.214) -13.829*** -3.029*** (0.551) (0.303) 754,121 754,121 -4.356*** -1.556*** (0.291) (0.102) -8.119*** -2.648*** Panel B: Education levels mother Vocational Upper secondary Upper secondary + ≥ 1 year College/University PhD Observations (0.473) (0.181) -10.015*** -2.854*** (0.381) (0.173) -12.324*** -3.242*** (0.508) (0.230) -14.553*** -3.443*** (1.059) (0.390) 754,121 754,121 Notes: Level indicates a dummy for each education level. The omitted education level is the lowest education level combined levels 1 and 2. Level of observation are sons of parents born between 1945 and 1955 in Sweden. Robust standard errors in parentheses, clustered by birth municipality. Signicance levels *** p<0.01, ** p<0.05, * p<0.1. All regressions include a full set of birth cohort dummies and birth municipality dummies. 50